High Signal Podcasts Evidence ledger
Method
Browse
← Back to evidence

Evidence receipt / belief

Published · transcript-backed

Speaker unverified: belief

3 Apr 2025 Latent Space The Creators of Model Context Protocol

“I mean, I think we've clearly done a lot of work. But I think we really need to be able to shift and like, you know, extract the information it cares about, because that's what that's exactly what they excel at.”

— Speaker unverified

Source trail

Everything needed to verify it.

Speaker
Speaker unverified
Attribution
Not verified from this transcript
Claim type
belief
Recorded
3 Apr 2025
Publisher
Latent Space

Transcript context

…Another question that I had when I was looking at some server implementations, the server builders kind of decide what data gets eventually returned, especially for tool calls. For example, the Google maps one, right? If you just look through it, they decide what, you know, attributes kind of get returned and the user can not override that if there's a missing one. That has always been my gripe with like SDKs in general, when people build like API wrapper SDKs. And then they miss one parameter that maybe it's new and then I can not use it. How do you guys think about that? And like, yeah, how much should the user be able to intervene in that versus just letting the server designer do all the work? I think we probably bear responsibility for the Google maps one, because I think that's one of the reference servers we've released. I mean, in general, for things like for tool results in particular, we've actually made the deliberate decision, at least thus far, for tool results to be not like sort of structured JSON data, not matching a schema, really, but as like a text or images or basically like messages that you would pass into the LLM directly. And so I guess the correlation that is, you really should just return a whole jumble of data and trust the LLM to like sort through it and see. I mean, I think we've clearly done a lot of work. But I think we really need to be able to shift and like, you know, extract the information it cares about, because that's what that's exactly what they excel at. And we really try to think about like, yeah, how to, you know, use LLMs to their full potential and not maybe over specify and then end up with something that doesn't scale as LLMs themselves get better and better. So really, yeah, I suppose what should be happening in this example server, which again, will request welcome. It would be great. It's like if all these result types were literally just passed through from the API that it's calling, and then the API would be able to pass through automatically. Thank you for joining us. It's a hard to sign decisions on where to draw the line.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.

Search evidence